experiments/launch.py

"""CPU-saturating batch launcher for geometric hypothesis experiments."""

from __future__ import annotations

import argparse
import json
import multiprocessing as mp
import os
from pathlib import Path
import sys
import traceback

WORKSPACE_ROOT = Path(__file__).resolve().parent.parent
if str(WORKSPACE_ROOT) not in sys.path:
    sys.path.insert(0, str(WORKSPACE_ROOT))

from encoder import config as encoder_config  # noqa: E402
from experiments import config  # noqa: E402


def available_cpu_count() -> int:
    """Use every CPU available through affinity unless explicitly overridden."""
    override = os.environ.get("VSI_CPU_WORKERS")
    if override is not None:
        count = int(override)
        if count < 1:
            raise ValueError("VSI_CPU_WORKERS must be positive")
        return count
    try:
        return max(1, len(os.sched_getaffinity(0)))
    except AttributeError:
        return max(1, os.cpu_count() or 1)


def configure_numerical_threads(count: int) -> None:
    """Give a worker its fair CPU share without nested-thread oversubscription."""
    value = str(max(1, count))
    for variable in (
        "OMP_NUM_THREADS",
        "MKL_NUM_THREADS",
        "OPENBLAS_NUM_THREADS",
        "NUMEXPR_NUM_THREADS",
    ):
        os.environ[variable] = value
    os.environ["VSI_KD_WORKERS"] = value


def scenes_with_existing_cache(depth, input_selection, tracking, frame_count):
    """Return scenes with both native caches, or an existing combined cache."""
    with encoder_config.JSONL.open(encoding="utf-8") as stream:
        scenes = list(
            dict.fromkeys(str(json.loads(line)["scene_name"]) for line in stream)
        )
    available = []
    for scene in scenes:
        combined = Path(
            encoder_config.cache_file(
                scene, depth, input_selection, tracking, frame_count
            )
        )
        da3 = Path(
            encoder_config.da3_cache_file(scene, depth, input_selection, frame_count)
        )
        sam3 = Path(
            encoder_config.sam3_cache_file(
                scene, input_selection, tracking, frame_count
            )
        )
        if combined.is_file() or (da3.is_file() and sam3.is_file()):
            available.append(scene)
    return available


def _worker(tasks, results, settings, threads_per_worker):
    configure_numerical_threads(threads_per_worker)
    import cv2

    cv2.setNumThreads(threads_per_worker)
    from experiments.run import run_scene

    while True:
        scene = tasks.get()
        if scene is None:
            return
        try:
            _, status, path = run_scene(scene=scene, **settings)
            results.put((scene, status, str(path)))
        except Exception:
            results.put((scene, "failed", traceback.format_exc()))


def launch(settings, scenes, workers=0):
    """Run scenes in parallel; zero workers means all available CPUs."""
    selected = list(scenes)
    if not selected:
        return {"built": 0, "loaded": 0, "failed": 0}
    cpu_count = available_cpu_count()
    worker_count = min(workers or cpu_count, len(selected))
    if worker_count < 1:
        raise ValueError("workers must be non-negative")
    threads_per_worker = max(1, cpu_count // worker_count)
    context = mp.get_context("spawn")
    tasks, results = context.Queue(), context.Queue()
    for scene in selected:
        tasks.put(scene)
    for _ in range(worker_count):
        tasks.put(None)
    processes = [
        context.Process(
            target=_worker,
            args=(tasks, results, settings, threads_per_worker),
        )
        for _ in range(worker_count)
    ]
    for process in processes:
        process.start()
    totals = {"built": 0, "loaded": 0, "failed": 0}
    for completed in range(1, len(selected) + 1):
        scene, status, detail = results.get()
        totals[status] += 1
        print(
            f"[{completed}/{len(selected)}] {scene}: {status} -> {detail}", flush=True
        )
    for process in processes:
        process.join()
    if totals["failed"]:
        raise RuntimeError(f"{totals['failed']} experiment scene(s) failed")
    return totals


def main() -> None:
    parser = argparse.ArgumentParser()
    parser.add_argument("--hypothesis", required=True)
    parser.add_argument(
        "--depth", default="metric", choices=encoder_config.DEPTH_VARIANTS
    )
    parser.add_argument(
        "--input",
        default="uniform",
        choices=encoder_config.INPUT_SELECTIONS,
        dest="input_selection",
    )
    parser.add_argument(
        "--tracking", default="tracking", choices=encoder_config.TRACKING_MODES
    )
    parser.add_argument("--frames", type=int, default=64)
    parser.add_argument(
        "--format",
        default="explicit",
        choices=config.SPATIAL_CODE_FORMATS,
        dest="spatial_code_format",
    )
    parser.add_argument(
        "--workers",
        type=int,
        default=0,
        help="worker processes; default 0 uses every available CPU",
    )
    parser.add_argument("--rebuild", action="store_true")
    parser.add_argument("--scene", action="append", dest="scenes")
    args = parser.parse_args()
    if args.frames < 1:
        parser.error("--frames must be positive")
    if args.workers < 0:
        parser.error("--workers cannot be negative")
    scenes = args.scenes or scenes_with_existing_cache(
        args.depth, args.input_selection, args.tracking, args.frames
    )
    settings = {
        "hypothesis": args.hypothesis,
        "depth": args.depth,
        "tracking": args.tracking,
        "input_selection": args.input_selection,
        "frame_count": args.frames,
        "rebuild": args.rebuild,
        "spatial_code_format": args.spatial_code_format,
    }
    totals = launch(settings, scenes, args.workers)
    print(
        f"DONE: {totals['built']} built, {totals['loaded']} loaded, "
        f"{totals['failed']} failed"
    )


if __name__ == "__main__":
    main()